# The 14T Mythos Leak Is Fake: Seeded Noise, Not Weights

> A 7.72 TB checkpoint sold as Anthropic's next Mythos-class model is seeded random noise. We regenerated its tensors from their names alone.

- Author: Nitish Garg, Founder & CEO, CellCog
- Published: 2026-11-01
- Canonical (HTML): https://cellcog.ai/blog/mythos-leak-hugging-face-fake/
- Section: Guides / Choosing a platform
- Publisher: CellCog (https://cellcog.ai), the AI employee platform. Blog index for agents: https://cellcog.ai/blog/llms.txt

## Key points

- On November 1, 2026 a 7.72 TB Hugging Face repository, burner22/numbat, spread on X as a mid-training checkpoint of Anthropic's next Mythos-class model.
- Its tensors are random numbers seeded from their own names: SHA-256 of the name, NumPy's PCG64 generator, a fixed scale.
- We reran the recipe at 20:18 UTC: 9 of 9 tensors in shard 1 matched exactly, 147,968 values; a wrong seed matched none.
- A Hugging Face user extended it to the large 4-bit matrices: signs matched on 162,246,496 of 162,246,496 values across 770 of 772 shards.
- Anthropic had not commented by 20:20 UTC. The real Mythos signal of the day was a hidden 5.5 label on its own Mythos page.

## At a glance

- **Is the Mythos leak on Hugging Face real?** No. The burner22/numbat weights regenerate bit for bit from their tensor names, so they were never trained.
- **Who proved it?** Hugging Face users in the repository's discussions on November 1, 2026; we reran their recipe and got 9 of 9 exact.
- **Should I download it?** No. It is 7.72 TB of generated numbers, and it ships custom code you would have to trust to run.

A 7.72 TB repository on Hugging Face, burner22/numbat, spread on November 1, 2026 as a mid-training checkpoint of Anthropic's next Mythos-class model, about 14 trillion parameters. It is fake. Its weights were never trained: they are random numbers seeded from each tensor's own name, and anyone can regenerate them bit for bit from a few kilobytes of header.

Hugging Face users posted that recipe within two hours of the leak post taking off on X. We reran it at 20:18 UTC: all 9 tensors we checked in the first shard matched exactly, 147,968 values, and a wrong seed matched none. Anthropic had not commented by 20:20 UTC, and its news page lists nothing about the repository.

## What the repository claims

![Animated diagram: a tensor name passes through SHA-256 and its first 8 bytes seed a gear, which fills a row of generated values that match the Hugging Face file cell for cell, stamped EXACT: 9 of 9 tensors, while a wrong seed gets 0 match](https://cellcog.ai/blog/media/mythos-leak-hugging-face-fake/seed.svg)
*Each tensor is regenerated from its own name, and the file matches it exactly*

The README describes a model of "~14T param, 114 layers", a mixture of experts with 1,224 experts (48 active per token plus one shared), stored in NVFP4, a 4-bit format. It calls numbat Anthropic's internal codename and points to a comment it says sat in the March 2026 Claude Code source leak: "Remove this section when we launch numbat." We have not verified that comment, and a codename, even a real one, says nothing about whose weights these are.

The repository was created on October 23, 2026 at 01:37 UTC by an account named burner22, and its files last changed on October 24. The README ends with "Mirrors welcome. Do not DM me asking for the source." By 20:03 UTC on November 1, Hugging Face's counter showed 5,846 downloads and 162 likes.

*Table: What the README claims vs what the files show*

| Claim | What the files show |
|---|---|
| A real Anthropic checkpoint | Norm and conv tensors regenerate from their names; no trained weights found |
| About 14 trillion parameters | 772 shards, 7.72 TB; the size is real, the content is generated |
| Internally consistent NVFP4 | True and beside the point: generated numbers quantize cleanly too |
| Separate output head | A commenter found it equals 0.891 times the embedding table |
| Runnable reference code | A commenter listed bugs: no KV cache, a no-op weight init, NaN on padded rows |
| Codename from a Claude Code leak | Not verified by us; a codename cannot make random weights trained |

## How the fake was found

On November 1 at 14:53 UTC @OxShamss posted the repository as one of the craziest AI leaks ever, while noting that nobody had verified the weights; the post passed 231,000 views by 20:03 UTC. A second account posted "ALL 7.72TB SECURED" a little after 17:00 UTC. On Hugging Face the checking had already started.

At 16:03 UTC a user posted a code review of the model files and called the release a well made fake. At 16:29 UTC a user named FloxUnder posted an audit: 43 complete tensors, 730,368 values, could be regenerated exactly from their names, with a short script that downloads about 90 KB. Within an hour others had reproduced it, one writing "I think that confirms it is fake." At 19:43 UTC a user named cekal extended the recipe to the large 4-bit matrices, which make up most of the file, and reported "162,246,496 out of 162,246,496 matches" on signs across 5,503 tensors from 770 of the 772 shards.

*Table: The numbat chain, dated (2026, UTC)*

| Time | Event | Source |
|---|---|---|
| Oct 23, 01:37 | Repository burner22/numbat created | Hugging Face API |
| Oct 24, 19:09 | Last file change | Hugging Face API |
| Nov 1, 14:53 | Posted on X as a Mythos-class leak | @OxShamss |
| Nov 1, 16:03 | Code review calls it a well made fake | Discussion #11 |
| Nov 1, 16:29 | 43 tensors regenerated from their names | Discussion #13 |
| Nov 1, 17:41 | Audit reproduced by another user | Discussion #13 |
| Nov 1, 19:43 | Recipe extended to the 4-bit matrices | Discussion #19 |
| Nov 1, 20:18 | Our rerun: 9 of 9 tensors exact | CellCog |

## We reran the check

The recipe is short. Read the first 8 bytes of shard 1 to get the header length, then the header itself, 16,344 bytes of JSON listing every tensor's name, shape and byte range. For each tensor, hash its name with SHA-256, take the first 8 bytes as a number, and use it to seed NumPy's PCG64 random generator. Draw normal random values in the tensor's shape, scale them (exp of 0.06 times the value for normalization weights, one sixth of the value for convolution kernels), round to bfloat16 and compare with the bytes on the server.

- 6 normalization tensors from layers 0 and 1: exact.
- 3 convolution kernels from layer 0: exact.
- 147,968 values in all, fetched with HTTP range requests.
- Control: the same tensor with seed 12345 did not match.

We downloaded only the header and those tensors, and ran none of the repository's code. A trained model cannot come out of a random generator whose only input is the tensor's name; these numbers carry no training at all.

## How to check a leaked checkpoint before you download it

- Read the header with a range request first. Never pull terabytes to answer a yes-or-no question.
- Test whether small tensors regenerate from a simple seed such as the name or the index. Trained norms and kernels never do.
- Do not load custom code from an anonymous repository. This one ships its own modeling file and asks for remote code to be trusted.
- Run the economics. One commenter wrote "This model is impossible to be real due to serving economics by any standard."
- Look for any first-party trace: a vendor page, a model card, a statement. Here there was none.

## Why someone would build it

Nobody knows. Commenters guessed a prank and, in one case, an attempt to move prices ahead of Anthropic's IPO; there is no evidence for either. What is clear is the care: matching configuration files, a README section arguing why readers should believe it, and 4-bit blocks that round-trip cleanly. One commenter noticed that "they even added noise to the initialized weights that could make a naive test assume they were actually trained."

## What this says about Mythos 5.5

Nothing. The repository predates the real signal of the day: Anthropic's own Mythos page showed a hidden Mythos 5.5 label in an October 31 archive copy, then read 5.1 again on November 1. That record, and what it means for the next Fable, is in our [Fable 5.5 and Mythos 5.5 tracker](https://cellcog.ai/blog/fable-5-5-release-date/). The last time a Mythos model was in the news for something real, it was [the AISI evaluation](https://cellcog.ai/blog/claude-mythos-5-aisi-incident/).

## Where CellCog stands

CellCog runs every tier on Claude Opus 5.5. Mythos models are available only to vetted organizations through Anthropic, and nothing in this repository changes what we route to.

## Sources

- Hugging Face, [burner22/numbat](https://huggingface.co/burner22/numbat), README, config and model API, revision 6b7f8fc5, read November 1, 2026, 20:03 UTC.
- Hugging Face discussions on the repository: [#11](https://huggingface.co/burner22/numbat/discussions/11), [#13](https://huggingface.co/burner22/numbat/discussions/13), [#14](https://huggingface.co/burner22/numbat/discussions/14) and [#19](https://huggingface.co/burner22/numbat/discussions/19), read November 1, 2026, 20:17 UTC.
- X: [@OxShamss, November 1](https://x.com/OxShamss/status/2116906417651126410). Read directly; time from the post id.
- NumPy, [PCG64 generator](https://numpy.org/doc/stable/reference/random/bit_generators/pcg64.html).
- CellCog, our rerun of the discussion #13 recipe on shard 1, November 1, 2026, 20:18 UTC.

## FAQ

**Is the 14 trillion parameter Mythos leak real?**

No. The tensors in burner22/numbat can be regenerated exactly from their names with a seeded random generator. Trained weights cannot be produced that way.

**What is numbat?**

The repository's README calls it Anthropic's internal codename and cites a comment from the March 2026 Claude Code source leak. We have not verified that comment, and Anthropic has not commented.

**How was the fake proven?**

Hash each tensor's name with SHA-256, seed NumPy's PCG64 generator with the first 8 bytes, draw normal values in the tensor's shape, scale and round to bfloat16. The result matches the file byte for byte.

**Is it safe to load the model?**

Do not run it. The weights are noise, the repository asks you to trust its own modeling code, and a commenter found bugs that make real inference impractical.

**Does this mean Mythos 5.5 is coming?**

This repository says nothing about it. Separately, Anthropic's Mythos page showed a hidden Mythos 5.5 label in an October 31 archive copy before reading 5.1 again.

## Related

- [Mythos 5.5 and Fable 5.5: The Label Anthropic Pulled](https://cellcog.ai/blog/fable-5-5-release-date/index.md)
- [Claude Mythos 5 AISI Incident: Why It Didn't Stop](https://cellcog.ai/blog/claude-mythos-5-aisi-incident/index.md)
- [Claude Opus 5.6 Release Date: The Routing Reports](https://cellcog.ai/blog/claude-opus-5-6-release-date/index.md)

## The AI employee for this read

[AI Software Engineer](https://cellcog.ai/ai-employees/ai-software-engineer): I built this page. For what it covers, hire an engineer: it works in your repo behind an approval gate, so nothing reaches your world unclassified.

---

Markdown alternate of https://cellcog.ai/blog/mythos-leak-hugging-face-fake/. Try CellCog free, no credit card needed: https://cellcog.ai/signup